Nadam Bound: A Novel Approach Of Subsuming Nesterov Momentum And Dynamic Bounding Into Adaptive Moment Estimation To Enhance The Detection Accuracy Of Deep Learning In Real World Steganalysis

Abstract

Within the realm of information security, steganography and stegan analysis are related concepts. The goal of steganalysis is to find hidden data in digital media. Steganographic techniques are always developing, leading to ongoing changes in steganalysis based on CNN. An optimizer in CNN tunes the weights while training, to reduce errors and boost the efficiency of the model. It is a key for any deep neural network learning to be both successful and efficient. Steganalysis based on deep learning is not an exception to the rule. Various deep network types usually require different optimizers, which must be selected through several experiments. To enhance the model's training speed effectively across deep networks, NAdamBound approach is introduced, a hybrid optimizer that combines Nadam's Nesterov momentum technique with Adabound's dynamic bounding mechanism. This could give rise to both the rapid convergence qualities from Nesterov momentum and the adjustable learning rate benefits.Extensive experiments conducted on Steganalysis of real-world dataset proved NAdamBound produced 94%, 92.2% and 93.8% accuracy against WOW, S-UNIWARD and HILL which surpasses its corresponding fellow optimizers in reducing loss and boosting accuracy.